Respiratory Feature Analysis for Micro-Expression Detection: Multimodal Approach on SAMM Dataset

Micro-expression detection faces significant challenges due to the fleeting and subtle nature of facial movements. Current approaches rely almost exclusively on visual features, yet recent findings suggest physiological signals might offer valuable complementary information for emotion recognition. This paper presents what we believe is the first systematic investigation of respiratory features for micro-expression recognition using the SAMM Long Videos dataset. We develop a non-contact respiratory signal extraction methodology and propose a 37-dimensional respiratory feature framework. In addition, we release the first open-source respiratory dataset for micro-expression research, providing synchronized respiratory signals and micro-expression annotations. Our experimental results show that respiratory features alone outperform traditional video-only baselines by 8.2% (macro-F1), while multimodal fusion of respiratory and video features yields further improvements of 17.0% over video-only methods. These findings indicate that respiratory patterns contain previously unexplored discriminative information for micro-expression detection, potentially establishing a new research direction in physiological affective computing.